Вы находитесь на странице: 1из 37

Journal of Computational Physics 221 (2007) 469505

www.elsevier.com/locate/jcp

A sharp interface method for incompressible two-phase ows


M. Sussman

*,1,

K.M. Smith, M.Y. Hussaini, M. Ohta, R. Zhi-Wei

Department of Mathematics, Florida State University, 208 Love Bldg., 002C Love Bldg., 487 DSL, Tallahassee, FL 32306-451, United States
Received 1 November 2005; received in revised form 8 June 2006; accepted 16 June 2006
Available online 27 July 2006

Abstract
We present a sharp interface method for computing incompressible immiscible two-phase ows. It couples the level-set
and volume-of-uid techniques and retains their advantages while overcoming their weaknesses. It is stable and robust
even for large density and viscosity ratios on the order of 1000 to 1. The numerical method is an extension of the second-order method presented by Sussman [M. Sussman, A second order coupled levelset and volume of uid method
for computing growth and collapse of vapor bubbles, Journal of Computational Physics 187 (2003) 110136] in which
the previous method treated the gas pressure as spatially constant and the present method treats the gas as a second incompressible uid. The new method yields solutions in the zero gas density limit which are comparable in accuracy to the
method in which the gas pressure was treated as spatially constant. This improvement in accuracy allows one to compute
accurate solutions on relatively coarse grids, thereby providing a speed-up over continuum or ghost-uid methods.
 2006 Elsevier Inc. All rights reserved.
MSC: 65M06; 76D05; 76T05
Keywords: Incompressible ow; Immiscible uids; NavierStokes equations; Multiphase ows; Numerical methods

1. Introduction
Ecient and accurate computation of incompressible two-phase ow problems has enormous value in
numerous scientic and industrial applications. Applications include ship hydrodynamics, viscoelastic free
surface ows, and liquid jets [10,14,15,48,36]. Current methods for the robustcomputation of immiscible
two-phase ows [51,50,9,53,32,26,20,18] are all essentially spatially rst-order accurate as the treatment of
the interfacial jump conditions constrains the overall accuracy to rst-order. Robustness is dened in terms
of the ability of a numerical method to stably handle wide ranges of physical and geometrical parameters.
We note that Hellenbrook et al. [21] developed a formally second-order level set method for two-phase ows,
but the applications did not include large density ratios, surface tension, or complex geometries. It is unlikely
that a straightforward application is possible to ow congurations with such wide parameter ranges. We also
note that Ye et al. [60] presented a second-order Cartesian grid/front tracking method for two-phase ows, but
*

Corresponding author. Tel.: +1 850 644 7194; fax: +1 850 644 4053.
E-mail address: sussman@math.fsu.edu (M. Sussman).
Work supported in part by the National Science Foundation under contracts DMS 0108672, U.S. Japan Cooperative Science 0242524.

0021-9991/$ - see front matter  2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.jcp.2006.06.020

470

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

their results did not include complex geometries. Yang and Prosperetti [57] presented a second-order boundary-tted tracking method for single-phase (free boundary problem) ows, but similarly as with Ye et al.
[60], their results did not include complex geometries.
Although the formal order of accuracy of continuum approaches [53,9,51,38,29] or ghost-uid approaches
[26,28] is second-order, numerical dissipation at the free surface reduces the order to rst-order. We propose a
new method which extends the functionality of the method discussed in [45] from single-phase (pressure
assumed spatially constant in the air) to multiphase (gas solution assumed incompressible). The resulting
matrix system(s) are symmetric, guaranteeing robustness of the method, and are capable of stably handling
wide parameter ranges (e.g. density ratio 1000:1, large Reynolds number) and geometries (e.g. topological
merging and breaking). The method is consistent in that it captures the limiting cases of zero gas density
and linear slip lines. Specically, the present method reduces to the single uid method [45] in the limit that
the gas density and gas viscosity approach zero (i.e. the numerical solution of the gas phase approaches the
condition of spatially constant pressure as the gas density approaches zero). The present method provides
additional functionality over single uid methods since one can accurately compute bubble entrainment, bubble formation, eect of wind on water, liquid jets, etc. Further, we demonstrate that the present method provides improved accuracy over existing two-uid methods for a given grid, and provides a speed-up over
existing methods for a given accuracy, as we can robustly compute ows on coarser meshes.
2. Governing equations
We consider the incompressible ows of two immiscible uids (such as liquid/liquid or liquid/gas), governed by the NavierStokes equations:
DU
r  pI 2lD qg^z
Dt
rU 0

where U is the velocity vector, q is the density, p is the pressure, l is the coecient of viscosity, g is the gravity,
I is the unit tensor, ^z is the unit vector in the vertical direction, and D is the deformation tensor dened by
T

rU rU
2
At the interface, C, separating the two uids, we have the normal continuity condition for velocity,
D

U  n  U L  n  U G  n 0
we also have the tangential continuity condition for velocity (if viscous eects are present),
U 0
and the jump condition for stress,
n  pI 2lD  n rj
where n is the unit normal to the interface, r is the coecient of surface tension and j is the local curvature.
Following the derivation in [12], we can rewrite the preceding governing equations in terms of the following
equations based on the level set function /. In other words, analytical solutions to the following level set equations are also solutions to the original governing NavierStokes equations for two-phase ow. Our resulting
numerical method will be based on the level set formulation.
If one denes the interface C as the zero level set of a smooth level set function, /, then the resulting equations are:
DU
r  pI 2lD qg^z  rjrH
Dt
rU 0
D/
0
Dt

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

q qL H / qG 1  H /
l lL H / lG 1  H /
r/
j/ r 
jr/j

1; / P 0
H /
0; / < 0

471

3
4

3. CLSVOF free surface representation


The free surface is represented by a coupled level set and volume-of-uid (CLSVOF) method [50]. In
addition to solving the level set equation (2), we also solve the following equation for the volume-of-uid function F,
DF
F t U  rF 0
Dt

(5) is equivalent to
F t r  UF r  UF
Since $ U = 0, we have
F t r  UF 0

At t = 0, F is initialized in each computational cell Xij,


Xij fx; yjxi 6 x 6 xi1

and

y j 6 y 6 y j1 g

as,
F ij

1
DxDy

H /x; y; 0 dx dy
Xij

Here, Dx = xi+1  xi and Dy = yi+1  yi.


The reasons why we couple the level set method to the volume-of-uid method are as follows:
 If one discretizes the level set equation (2), even in conservation form, the volume enclosed by the zero level
set will not be conserved. This problem has been addressed by implementing global mass xes [12], augmenting the level set equation by advecting massless particles [16,24], implementing adaptive mesh renement techniques [43,47], and by implementing high order spectral methods [49,30]. In this paper, we
preserve mass by coupling the level set method to the volume-of-uid method [8,50,58]; eectively, by coupling the two, we are implementing a local mass x instead of a global mass x. The volume-of-uid
function F is used to correct the mass enclosed by the zero level set of / during the level set redistancing
step (see Fig. 1).
 If one uses only the volume-of-uid function F to represent the interface separating air and water, then one
must be able to accurately extract the normal and curvature from F. Also, small pieces of volume might
separate from the free surface which can pollute the solution for the velocity. Modern volume-of-uid
methods have addressed these problems [33,22,18] using second-order slope reconstruction techniques
and calculating the curvature either from a height fraction or from a temporary level set function.
The level set function in our implementation is used for calculating the interface normal and is used for
calculating density and viscosity used by the NavierStokes equation. The level set function is not used
for calculating the interface curvature; instead we use the volume-of-uid function.
To clarify what information we extract from the level set function /, and what information we extract from
the volume-of-uid function F, we have:

472

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

Closest point.

(i,j+1)

(i+1,j1)
Fig. 1. After each time step, the level set function / is reinitialized as the closest distance to the piecewise linear reconstructed interface.
r/
. In this diagram, the shaded area fraction is Fi, j+1,
The linear reconstruction encloses the volume given by F with its slope given by n jr/j
the distance from point xi+1, j1 to the closest point becomes the new value of /i+1, j1.

 The normals used in the volume-of-uid reconstruction step are determined from the level set function (see
e.g. Fig. 2).
 The height fraction (see Section 5.3) and velocity extrapolation calculations (see Section 5.6) both depend
on the level set function. Therefore, cells in which F is very close to either 0 or 1 will not directly eect the
accuracy of the solution to the momentum equations.
 The volume fractions are used, together with the slopes from the level set function, to construct a volumepreserving distance function along with providing closest point information to the zero level set (see
Fig. 1).

(i,j+1/2)

(i+1/2,j)

(i1/2,j)

(i,j1/2)

x
Fig. 2. In order to calculate the volume-of-uid ux, Fi+1/2, j, one must rst nd the linear coupled level set and volume-of-uid
reconstructed interface. The ux across a face becomes the volume fraction of overall volume that is advected across a face. For this
area
illustration, F i1=2;j shaded
.
ui1=2;j DtDy

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

473

 The volume fractions are used to express the interfacial curvature to second-order accuracy (see Section
A.2). We do not use the level set function for nding the curvature because our level set reinitialization step
is only second-order accurate; the curvature as computed from the level set function will not provide the
second-order accuracy that is provided directly from the volume fractions.
We observe that there are possibly more accurate representations of the interface [49,41,16,4,3,40]. However, it must be noted that the accuracy of the computations is limited by the order of accuracy of the treatment of the interfacial boundary conditions and not by the accuracy of the interface representation. Even if
the interface representation is exact, if the velocity used to advance the interface is low order accurate, then the
overall accuracy is constrained by the accuracy at which the velocity eld (specically, the velocity eld at the
interface) is computed. Our results in Sections 6.3 and 6.4 support this hypothesis. We demonstrate secondorder accuracy for interfacial ows in which only rst-order methods have been previously applied. We also
show that we conserve mass to a fraction of a percent in our computations (e.g. largest mass uctuation on
coarsest grid in Section 6.3 was 0.08%).
When implementing the CLSVOF method, the discrete level set function /nij and discrete volume fraction
function F nij are located at cell centers. The motion of the free surface is determined by the face centered velocities, ui+1/2,j and vi,j+1/2, which are derived from the momentum equation. A scalar quantity with the subscript
ij implies that the quantity lives at the cell center (xi, yj),
xi xl0 i 1=2Dx
y j y l0 j 1=2Dy
A scalar quantity with the subscript i + 1/2, j implies that the quantity lives at the right face center of cell ij,
xi1=2 xl0 i 1Dx
y j y l0 j 1=2Dy
A scalar quantity with the subscript i, j + 1/2 implies that the quantity lives at the top face center of cell ij,
xi xl0 i 1=2Dx
y j1=2 y l0 j 1Dy
A vector quantity with the subscript i + 1/2, j implies that the rst component lives at the right face center of a
cell, (xi+1/2,yj), and that the second component lives at the top face center of a cell, (xi, yj+1/2). A diagram illustrating where our discrete variables live is shown in Fig. 3.
The discrete face centered velocity eld is assumed to satisfy the discrete continuity condition at every point
in the liquid (/ij P 0):
ui1;j  ui12;j vi;j12  vi;j12

0
7
Div Uij 2
Dx
Dy
To integrate the solution for both the level set function / and the volume-of-uid function F, we rst simultaneously solve (6) and (2). Then, we reinitialize / by constructing a distance function that shares the same
enclosed volume as determined from F, and the same slopes as determined from /.
Both the level set equation and the volume-of-uid equation are discretized in time using second-order
Strang splitting [44] where for one time step we sweep in the x direction then the y direction, then for
the next time step, we sweep in the y direction, then the x direction. Assuming that the advective velocity
is independent of time, this procedure is equivalent to solving for the x direction terms for Dt time, solving
y direction terms for 2Dt time, then solving for the x direction terms again for Dt time.
The spatial operators are split, where one alternates between sweeping in the x direction:
F ij  F nij ui1=2;j F ni1=2;j  ui1=2;j F ni1=2;j
ui1=2;j  ui1=2;j

F ij
Dt
Dx
Dx
/ij  /nij ui1=2;j /ni1=2;j  ui1=2;j /ni1=2;j
ui1=2;j  ui1=2;j

/ij
Dt
Dx
Dx

474

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

(i,j+1)
(i,j+1/2)

(i1,j)

(i1/2,j) (i,j) (i+1/2,j)

(i+1,j)

(i,j1/2)
(i,j1)

Fig. 3. Cell centered quantities, /ij, Fij, pij, live at the cell locations (i, j), (i + 1, j), (i, j + 1), etc. The horizontal MAC velocity, ui+1/2, j, lives
at the vertical face centroids, (i  1/2, j), (i + 1/2, j), etc. The vertical MAC velocity, vi, j+1/2 lives at the horizontal face centroids, (i, j  1/2),
(i, j + 1/2), etc.

and in the y direction:


F ijn1  F ij vi;j1=2 F i;j1=2  vi;j1=2 F i;j1=2
vi;j1=2  vi;j1=2

F ij
Dt
Dy
Dy


n1

/ij  /ij vi;j1=2 /i;j1=2  vi;j1=2 /i;j1=2
vi;j1=2  vi;j1=2

/ij
Dt
Dy
Dy

The volume-of-uid uxes, Fi+1/2,j and Fi,j+1/2, are calculated as the fraction of liquid uid to the overall uid
that is advected across a given cell face during a timestep (see Fig. 2). The level set uxes, /i+1/2,j and /i,j+1/2
are calculated by extrapolating the level set function in space and time to get a time-centered ux at given cell
faces. Details are presented in [50,45].
If we add 8 to 9, then we have,
F n1
 F nij ui1=2;j F ni1=2;j  ui1=2;j F ni1=2;j vi;j1=2 F i;j1=2  vi;j1=2 F i;j1=2
ij

Dt
Dx
Dy


u

u
v

v
i1=2;j
i1=2;j
i;j1=2
i;j1=2

F ij
Dx
Dy

10

If the right hand side of (10) is zero, then F shall be conserved since the left hand side of (10) is written in
conservation form. In other words, if the discrete divergence free condition (7) is satised, then we have mass
conservation. A key distinction between the two-phase algorithm we present here and previous sharp interface
methods is that solutions derived from our method will approach the solutions of the corresponding one-phase
method in the limit that the vapor is assumed to have uniform pressure. In order to achieve this goal, we
implement a liquid velocity extrapolation procedure in which we extrapolate the liquid velocities into the
gas (therefore, we shall store two separate velocity elds). The extrapolated liquid velocity may not satisfy
(7) in vapor cells (/ij < 0). In order to maintain conservation of F, we have the additional step,


vi;j1=2  vi;j1=2
n1
n1
 ui1=2;j  ui1=2;j

F ij F ij  DtF ij
11
Dx
Dy
The resulting advection procedure for F now becomes,
F n1
 F nij ui1=2;j F ni1=2;j  ui1=2;j F ni1=2;j vi;j1=2 F i;j1=2  vi;j1=2 F i;j1=2
ij

0
Dt
Dx
Dy

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

475

We remark that in [45], we required that Eq. (7) hold in both liquid cells (/ij P 0) and extrapolated cells; this
requirement necessitated an extrapolation projection step. In this work, we relax this condition and instead
use (11).
4. Temporal discretization: CrankNicolson/TVD RungeKutta, projection method
Our temporal discretization procedure for approximating Eq. (1) is based on a combination of the Crank
Nicolson projection procedure (see e.g. [5,6]) for the viscous terms and the second-order TVD preserving
RungeKutta procedure [42] for the nonlinear advective terms.
Our method follows loosely the outline below:
Sweep 1
U n1;0 U n DtF U n Dt

GU n GU n1;0
 Dt Grad P n
2

12

Sweep 2
U n1;1 U n DtF U n1;0 Dt
U n1

GU n GU n1;1
 Dt Grad P n1
2

U n1;0 U n1;1
2

13

where F corresponds to the nonlinear advective terms, G corresponds to the viscous terms and Grad P
corresponds to the pressure gradient term. To be more specic, we describe one sweep of our method below.
Prior to each timestep we are given a liquid velocity, UL,n, and a total velocity Un. The main distinction
between our method and previous sharp interface methods is that we store UL,n in addition to storing Un.
In a given time step, immediately after solving for Un+1, we construct UL,n+1,
(
/ij P 0 or /i1;j P 0
ui1=2;j
L
ui1=2;j
extrapolate
otherwise
ui1=2;j
In other words, UL corresponds to U except on gas faces, where we replace the gas velocity in UL with the
extrapolated liquid velocity. UL is then used to calculate the nonlinear advective terms in the liquid, and also
used to advance the free surface.
Prior to each time step, we are also given a live pressure gradient,

n
rp rjrH
n
Grad P 
q
a level set function, /n, and a volume-of-uid function, Fn. The live pressure gradient, level set function, and
volume-of-uid function are stored at cell centers. The velocity is stored at both cell centers and face-centers.
As previously noted in Section 3, the subscript ij refers to the center of a computational cell, the subscript
i + 1/2, j refers to the right face center of a cell, and the subscript i, j + 1/2 refers to the top face center of a
cell. A vector quantity with the subscript i + 1/2, j implies that the rst component lives at the right face center
of a cell and the second component lives at the top face center of a cell.
A representative outline of one sweep of our (two-phase) method follows.
Step 1. CLSVOF [50,45] interface advection:
/ijn1 /nij  DtU L  r/ij
F n1
F nij  DtU L  rF ij
ij

476

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

Step 2. Calculate (cell centered) advective force terms:


ALij U L  rU L nij
Aij U  rUnij
Details for the calculation of these terms are presented in Section 5.1 below.
Step 3. Calculate (cell centered, semi-implicit) viscous forces:
(
/ij P 0
U L;n
ij
n
U ij
U nij /ij < 0
(
ALij /ij P 0
Aij
Aij /ij < 0
(
qL /ij P 0
qij
qG /ij < 0

14


n
U ij  U nij
1 Lij Lij
Aij g^z  Grad P nij
qij
Dt
2

The discrete operator L is a second-order approximation to $ 2lD (see Section 5.4).


Step 4. Interpolate cell centered forces to face centered forces:
1
ALi1=2;j ALi1;j ALi;j
2
1
Ai1=2;j Ai1;j Ai;j
2(
ALi1=2;j /ij P 0 or /i1;j P 0
Ai1=2;j
Ai1=2;j otherwise
1
Li1=2;j Lij Lnij Li1;j Lni1;j
4
(
/ij P 0 or /i1;j P 0
U L;n
i1=2;j
U ni1=2;j
n
U i1=2;j otherwise
V i1=2;j

U ni1=2;j

15

!


2
rjrH
Dt Ai1=2;j
Li1=2;j 
g^z
qi1;j qi;j
q
i1=2;j

See Section 5.2 for steps to discretize the surface tension force q1 rjrH .
Step 5. Implicit pressure projection step:
rp
rV
r
q
 
rp
n1
U i1=2;j
V
q i1=2;j

16

Section 5.5 provides the spatial discretization associated with the implicit pressure projection step. We
solve the resulting linear system using the multigrid preconditioned conjugate gradient method
(MGPCG) [52].
L;n1
n1
Step 6. Liquid velocity extrapolation; assign U L;n1
i1=2;j U i1=2;j and then extrapolate U i1=2;j into the gas region
(see Section 5.6).
Step 7. Interpolate face centered velocity to cell centered velocity:
1
L;n1
U ijL;n1 U L;n1
U i1=2;j

2 i1=2;j
1 n1
n1
U ijn1 U i1=2;j
U i1=2;j

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

477

Step 8. Update the cell centered live pressure gradient term,

Grad P n1
ij

U ij  U ijn1
Grad P nij
Dt

Often in this paper, we shall compare the two-phase algorithm just described, to the corresponding
one-phase algorithm. So, in the appendix (Section A.1) we describe the one-phase equations and
algorithm.
Remarks:
 At the very rst time step, we initialize
Grad P 0  Grad P 00  0
then we do ve iterations of the CrankNicolson/RungeKutta procedure ((12) and (13)) in order to initialize an appropriate cell centered pressure gradient,
Grad P 0 Grad P 05
We have found empirically that the cell centered pressure gradient term suciently converges after 5
sweeps. For example, for the very rst time step for the problem of the break-up of a cylindrical jet due
to surface tension (Section 6.5), the relative error in the magnitude of Grad P0(5) is 0.0008.
 If Step 6 (velocity extrapolation) is ignored, then our method corresponds in spirit to the sharp interface
ghost-uid approach described in [26,28]. This is because, without velocity extrapolation, UL = U. In this
case, when UL = U, the main dierence separating our approach from previous sharp interface methods
[26,28] is that we treat the viscosity jump conditions implicitly; therefore we have no time step constraints
associated with viscosity. We shall label this method, where liquid velocity extrapolation is ignored, as the
semi-implicit ghost-uid method.
 If Step 6 (velocity extrapolation) is not ignored, then our method has the property that, for the limiting
case of zero gas density and zero gas viscosity, our two-phase method is discretely equivalent to the secondorder one-phase approach [45] in which gas pressure is treated as spatially uniform; Section A.1 gives a
review of the one-phase approach.
5. Spatial discretization
5.1. Nonlinear advective terms
The term,
U  rUij
is discretized as
0 u
1


u

u
u
uij i1=2;jDx i1=2;j vij i;j1=2Dy i;j1=2
@
A
v
v
v
v
uij i1=2;jDx i1=2;j vij i;j1=2Dy i;j1=2
i;j1=2 and vi;j1=2 are constructed from the cell centered velocity eld Uij using
i1=2;j , vi1=2;j , u
The quantities u
upwind and slope-limited dierencing; e.g.
(
if vi;j1=2 > 0
uij 12 uy;ij

ui;j1=2
ui;j1  12 uy;i;j1 if vi;j1=2 < 0
The slopes uy,ij are computed using second-order Van Leer slope limiting [54],

S min2jui;j1  ui;j j; 2jui;j  ui;j1 j; 12 jui;j1  ui;j1 j if s > 0
uy;ij
0
otherwise

478

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

where
S signui;j1  ui;j1
and
s ui;j1  ui;j ui;j  ui;j1
5.2. Surface tension force
In this section, we describe the discretization of the face centered surface tension term,
rji1=2;j rH i1=2;j
qi1=2;j
which is found in Eq. (15).
The discretization of the face centered surface tension term at the face center, (i + 1/2,j), is written as,
H /

H /ij

rji1=2;j i1;jDx
qi1=2;j

17

where

H /

/P0

/<0

and
qi1=2;j qL hi1=2;j qG 1  hi1=2;j
The discretization of the height fraction, hi+1/2,j, is given in Section 5.3 (also see [19,28]).
The curvature ji+1/2,j is computed with second-order accuracy directly from the volume fractions as
described in Section A.2.
Our treatment of surface tension can be approached from two dierent perspectives: (1) the surface tension
term is derived in order to enforce the pressure jump condition as with the ghost-uid approach [26,28]; (2) the
inclusion of the surface tension term (17) as a force term in the momentum equation (15) is equivalent to prescribing the second-order Dirichlet pressure condition of surface tension (39) that would occur if the gas pressure was treated as spatially constant and the gas was assumed to be a void (qG = 0).
5.2.1. Ghost-uid perspective for the surface tension term
Here we give the ghost-uid [26] derivation of the surface tension term for the inviscid Eulers equations.
Without loss of generality, we consider a free surface that is vertically oriented between cells (i, j) and (i + 1, j),
at the location (xi+1  hDx, yj), with liquid on the right and gas on the left (see Fig. 4). At the face separating
cells (i, j) and (i + 1, j), the updated velocity is given by,
n1
ui1=2;j
 ui1=2;j

Dt
n1
ui1=2;j
 ui1=2;j
Dt

rpL
qL

rpG
qG

The continuity condition requires that,


rpL
rpG
n G n
L
q
q

18

and the pressure jump condition requires that,


pLI  pG
I rj

19

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

Gas

(i,j)

Liquid

(i+1,j)

(i+1/2,j)

479

Fig. 4. Ghost-uid treatment for vertical interface with liquid on the right and gas on the left. pi+1,j is the liquid pressure at cell (i + 1, j)
and pi,j is the gas pressure at cell (i, j). pLI is the liquid pressure on interface I and pG
I is the gas pressure on interface I. hDx is the distance
from the interface, I, to the liquid cell (i + 1, j).
L
where the gas and liquid pressure on the free surface are pG
I and p I , respectively. As a result of discretizing (18),
one has,
G
pLi1;j  pLI
pG
I  p i;j

qL hDx
qG 1  hDx

20

After one solves (19) and (20) for pLI and pG


I , and substitutes the results back into the liquid and gas pressure
gradients, one has,
n1
ui1=2;j
 ui1=2;j

Dt

pi1;j pi;j
Dx

qi1=2

 rjI

H i1;j H i;j
Dx

qi1=2

The surface tension term here is equivalent to (17). Without liquid velocity extrapolation, or if the gas density
is not negligible, then this discretization will be rst-order accurate since we had assumed in our derivation
that the free surface was oriented either vertically or horizontally. Now suppose that we keep the liquid velocity separate from the gas velocity during the calculation of the nonlinear advection terms, and also suppose
that the gas density is negligible, then one can relate the surface tension term to the Dirichlet boundary condition that one would impose for a one-phase method.
5.2.2. One-phase perspective for the surface tension
Here we give the one-phase [19,45] derivation of the surface tension term for the inviscid Euler equations.
In contrast to the two-phase case where there were two boundary conditions at the interface, there is only one
condition on pressure at the interface for the one-phase free boundary problem:
pLI pG
I  rj
The momentum equation in the liquid phase is,
uL;n1  uL;
rpL
 L
Dt
q
Suppose we are considering a free surface that passes between cells (i, j) and (i + 1, j), at the location
(xi+1  hDx, yj), with liquid on the right and gas on the left (see Fig. 4). Also, denote the gas and liquid presL
sure on the free surface as pG
I and p I , respectively. Then, discretely, one has,
p

p

i1
i
pL  pG
rpL
I  rjI
 i1
 DxL  rjI
L
L
q
hq
hDxq

H i1;j H i;j
Dx
hqL

480

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

This latter formulation is equivalent to the former when qG = 0 and we assume that gas pressure is spatially
uniform. In other words, our treatment for surface tension corresponds to the treatment in a second-order
single-phase approach (see Section A.1 or [45]). As mentioned by [19], this specication of the pressure
boundary condition is second-order accurate; as opposed to the ghost-uid perspective, we do not have
to make assumptions regarding the orientation of the interface in order to get second-order.
5.3. Height fraction
The height fraction hi+1/2,j [19,28,45] gives the one-dimensional fraction of water between cells (i, j) and
(i + 1, j). Fig. 5 gives an illustration of the height fraction. The mixed face-centered density is expressed in
terms of the height fraction. The height fraction hi+1/2,j is derived from the level set function as follows:
8
1
/i1;j P 0 and /i;j P 0
>
>
<
/i1;j < 0 and /i;j < 0
hi1=2;j / 0

>
/
/
>
: i1;j i;j otherwise
j/i1;j jj/i;j j

The + superscript stands for the positive part: i.e. a+ max(a, 0).
5.4. Semi-implicit viscous solve
An important property of our sharp-interface treatment for the viscous force terms is that resulting solutions of our two-phase algorithm approach solutions of the one-phase algorithm in the limit of zero gas density and zero gas viscosity (i.e. in the limit, in which the gas pressure is treated as spatially uniform).
The viscous force terms, Lij and Lnij , appear in the discretized NavierStokes equations as shown below,

n
U ij  U nij
1 Lij Lij
Aij g^z  Grad P nij
qij
Dt
2

21

L is a second-order discretization of the viscous force term, $ 2lD. In two dimensions, the rate of deformation tensor D is given by,

Gas

=0
i-1/2,j+1

i+1/2,j+1=5/8

i+1/2,j=1
+d

i,j=+d
i-1/2,j =5/8

Water
Fig. 5. Illustration of the face-centered height fraction, hi+1/2, j, hi, j+1/2.

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505


D

ux

uy vx =2

uy vx =2

vy

481

In previous work [50], we found the rate of deformation tensor D at cell faces and used a nite volume discretization to approximate $ 2lD. In other words, in previous work we had,
0 2l
1
li;j1=2 uy vx i;j1=2 li;j1=2 uy vx i;j1=2
i1=2;j ux i1=2;j 2li1=2;j ux i1=2;j

Dx
Dy
A
r  2lDij  @ l
2li;j1=2 vy i;j1=2 2li;j1=2 vy i;j1=2
i1=2;j uy vx i1=2;j li1=2;j uy vx i1=2;j

Dx
Dy
For a sharp interface method based on the nite volume discretization, the viscosity at a face is given by
[26,28],
8
l
hi1=2;j 1
>
>
> L
>
< lG
hi1=2;j 0
li1=2;j
0
l
>
G 0 and 0 < hi1=2;j < 1
>
>
>
l
l
G
L
:
otherwise
lG hi1=2;j lL 1hi1=2;j
Unfortunately, with the above discretization for the viscosity term, the two-phase method does not correspond to the single-phase method (Section A.1) when lG = 0. This is because velocities in gas cells could be
accidentally included in the discretization of the coupling terms in liquid cells, even if lG = 0. Fig. 6 gives an
illustration of how gas velocities can be accidentally included in the discretization of the coupling terms (the
term (lvx)y in the rst equation and the term (luy)x in the second). Therefore, we use the following node
based discretization instead of the preceding nite volume discretization:


 1
0
oluy vx
o2lux

ox
oy
ij
ij C
B


 A
r  2lDij @ 
oluy vx
o2lvy

ox
oy
ij

ij

where



o2lux
 2li1=2;j1=2 ux i1=2;j1=2  2li1=2;j1=2 ux i1=2;j1=2
ox
ij
2li1=2;j1=2 ux i1=2;j1=2  2li1=2;j1=2 ux i1=2;j1=2

.
2Dx

(i+1/2,j+1/2)

Gas
=0

=0

Liquid
Fig. 6. Illustration of how the gas velocity at cell (i + 1, j + 1) is inadvertently included in the calculation of the coupling terms when
lG = 0 and when the viscosity coecient is given at the cell faces, li+1/2,j, etc. The + and  signs refer to the sign of the level set function.
In this scenario, all the face centered coecients are equal to the liquid viscosity coecient. If the viscosity coecient is given at the nodes,
and if lG = 0, then li+1/2,j+1/2 = 0 and the gas velocity at (i + 1, j + 1) will not be included in the calculation of the viscous coupling terms.

482

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

oluy vx
oy


 li1=2;j1=2 uy vx i1=2;j1=2  li1=2;j1=2 uy vx i1=2;j1=2

ij

li1=2;j1=2 uy vx i1=2;j1=2  li1=2;j1=2 uy vx i1=2;j1=2





oluy vx
 li1=2;j1=2 uy vx i1=2;j1=2  li1=2;j1=2 uy vx i1=2;j1=2
ox
ij


o2lvy
oy

li1=2;j1=2 uy vx i1=2;j1=2  li1=2;j1=2 uy vx i1=2;j1=2

.

.

2Dy

2Dx


 2li1=2;j1=2 vy i1=2;j1=2  2li1=2;j1=2 vy i1=2;j1=2

ij

2li1=2;j1=2 vy i1=2;j1=2  2li1=2;j1=2 vy i1=2;j1=2


The viscosity at a node is given by
8
lL
>
>
>
>
< lG
li1=2;j1=2
0
>
>
>
>
lG lL
:

lG hi1=2;j1=2 lL 1hi1=2;j1=2

.

2Dy

hi1=2;j1=2 1
hi1=2;j1=2 0
lG 0 and 0 < hi1=2;j1=2 < 1
otherwise

where hi+1/2,j+1/2 is a node fraction dened as,


8
1
/i1;j P 0; /i;j P 0; /i;j1 P 0 and /i1;j1 P 0
>
>
<
/i1;j < 0; /i;j < 0; /i;j1 < 0 and /i1;j1 < 0
hi1=2;j1=2 / 0

>
/
/
/
/
>
: i1;j i;j1 i;j i1;j1
otherwise
j/i1;j jj/i;j1 jj/i;j jj/i1;j1 j
The + superscript stands for the positive part: i.e. a+ max(a,0).
The components of the deformation tensor, e.g. (ux)i+1/2,j+1/2, are calculated using standard central dierencing, i.e.
ui1;j1 ui1;j  ui;j1  ui;j
ux i1=2;j1=2
2Dx
The resulting linear system (21) for U* is solved using the standard multigrid method.
Remarks:
 Our discretization of the viscous forces are second-order accurate away from the gasliquid interface, but
only rst-order accurate at the gasliquid interface. We observe rst-order accuracy whether we are implementing our semi-implicit viscous solver as a part of the single-phase algorithm or as a part of the twophase algorithm. Only in places where the free surface is aligned exactly with grid boundaries would our
discretization be second-order accurate.
 Our proposed discretization of the node fraction, hi+1/2, j+1/2, is not necessarily the only possible choice. The
critical property that any discretization technique for the node fraction must have, is that hi+1/2, j+1/2 < 1 if
any of the surrounding level set values are negative.
5.5. Projection step
In this section, we provide the pertinent details for the discretization of the projection step found in Eq.
(16),
rp
r
rV
22
q
rp
UV
23
q

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

483

Eqs. (22) and (23) are discretized as


Div

Grad P
Div V
q

24

and
Grad P
q
respectively. Div is the discrete divergence operator dened by
U V

ui1=2;j  ui1=2;j vi;j1=2  vi;j1=2

;
Dx
Dy

25

and Grad represents the discrete gradient operator,


pi1;j  pi;j
Grad pi1=2;j
Dx
pi;j1  pi;j
Grad pi;j1=2
Dy

26

Div Vij

27

so that (24) becomes,


pi1;j pij
qi1=2;j

pij pi1;j
qi1=2;j

Dx2

pi;j1 pij
qi;j1=2

pij pi;j1
qi;j1=2

Dy 2

Div V

The face centered density is dened by


qi1=2;j qL hi1=2;j qG 1  hi1=2;j

28

where the discretization of the height fraction, hi+1/2, j, is given in Section 5.3.
At impenetrable boundaries, we give the Neumann boundary condition,
rp  n 0
and we also modify V to satisfy,
V n0
At outow boundaries, we give the Dirichlet boundary condition,
p0
i.e. if the top wall is outow, then we have pi;jhi 1 pi;jhi .
The resulting discretized pressure equation, (24), is solved for p using the multigrid preconditioned conjugate gradient method [52].
Remark:
 In the limit as qG approaches zero, one recovers the second-order projection step described in [45]. In other
words, in the limit of zero gas density, one recovers the second-order discretization of Dirichlet boundary
conditions at the free surface. The discretization, using the height fractions hi+1/2,j, corresponds to the second-order method described by [19] (in the zero gas density limit).
 By storing the velocity eld at the cell faces and the pressure at the cell centers, we avoid the checkerboard instability while maintaining a discretely divergence free velocity eld.
 We construct a temporary cell centered velocity eld for calculating the advection and diusion terms. Since
at each timestep we interpolate the advective and diusive forces from cell centers to cell faces in preparation for the next projection step, we avoid unnecessary numerical diusion that would occur if we had interpolated the velocity itself from cell centers to cell faces.

484

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

5.6. Extrapolation of MAC velocities


The liquid velocity uLi1=2;j is extended in a small narrow band about the zero level set of the level set function /. Extension velocities are needed on gas faces (i + 1/2,j) that satisfy /i,j < 0 and /i+1,j < 0. We describe
the initialization of uLi1=2;j below; the case for vLi;j1=2 follows similarly. The extension procedure is very similar
to that described in [45], except that (1) we choose an alternate, more stable, method for constructing our second-order linear interpolant and (2) we do not project the extended velocity eld; in lieu of projecting the
extended velocity eld, we instead discretize the volume of uid Eq. (6) in conservation form.
The steps for our liquid velocity extrapolation procedure are:
1. For each point where /i+1, j < 0 and /i, j < 0 and (1/2)(/ij + /i+1, j) > KDx, we already know the corresponding closest point on the interface xclosest, i+1/2, j (1/2)(xclosest, ij + xclosest, i+1, j). The closest point on
the interface has already been calculated during the CLSVOF reinitialization step (details found in [45], also
see Fig. 1) since the distance at a gas cell xij is,
d jxij  xclosest;ij j
2. Construct a 7 7 stencil for ui+1/2, j about the point xclosest, i+1/2, j. A point xi0 1=2;j0 in the stencil is tagged as
valid if /i0 ;j0 P 0 or /i0 1;j0 P 0. A diagram of how this 7 7 stencil is created for extending the horizontal
velocity uextend
i1=2;j is shown in Fig. 7. Please see Fig. 8 for a diagram portraying the 7 7 stencil used for constructing the vertical extension velocities vextend
i;j1=2 .
3. Determine the valid cell (icrit + 1/2, jcrit) in the 7 7 stencil that is closest to xclosest, i+1/2, j.
4. Determine the slopes Dxu and Dyu. In the x direction, investigate the forward dierences,
Dx u ui0 3=2;jcrit  ui0 1=2;jcrit
where (i 0 + 3/2, jcrit) and (i 0 + 1/2, jcrit) are valid cells in the 7 7 stencil. In the y direction, investigate the
forward dierences,
Dy u uicrit 1=2;j0 1  uicrit 1=2;j0
where (icrit + 1/2, j 0 + 1) and (icrit + 1/2, j 0 ) are valid cells in the 7 7 stencil.If any of the dierences change
sign in the x(y) direction, then the slope, Dxu(Dyu) is zero, otherwise the slope is taken to be the quantity
Dxu(Dyu) that has the minimum magnitude.

phi<0
F=0

phi<0
F=2/5

phi>0
F=1

Fig. 7. Diagram hi-lighting the valid points in the 7 7 stencil used for constructing the horizontal extension velocities.

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

485

phi<0
F=0

phi<0
F=2/5

phi>0
F=1

Fig. 8. Diagram hi-lighting the valid points in the 7 7 stencil used for constructing the vertical extension velocities.

5. Construct
crit
uextend
Dy uj  jcrit uicrit 1=2;jcrit
i1=2;j Dx ui  i

5.7. Timestep
The timestep Dt at time tn is determined by restrictions due to the CFL condition, surface tension, and
gravity:
0
1
r
L
1
2Dx
B1 Dx 1 q
C
qA
;
Dx3=2 ;
Dt < min @
i;j
2 jU n j 2 8pr
2 n
2
ju j jun j 4gDx
The stability condition regarding gravity was determined heuristically in which we have the inequality,
u DtgDt < Dx
The stability condition for surface tension is taken from [9,18]. Other references regarding stability conditions
for incompressible ow are [2,31].
6. Results
In this section we test the accuracy of our numerical algorithm. In a few cases, we shall compare our sharp
interface approach to the semi-implicit ghost-uid approach. Also, we shall compare our two-phase sharp
interface approach to our one-phase sharp interface method. In cases where the exact solution is unknown,
we calculate the error by comparing the solutions on successively rened grids. The error in interface position
is measured as
XZ
Einterface
jH /f  H /c j dx
29
ij

Xij

where /f and /c correspond to the solutions using the ne resolution and coarse resolution grids, respectively.

486

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

The average error in liquid velocity is measured as (for 3d-axisymmetric problems),

X q
2
2
avg
ELiquid
uf;ij  uc;ij vf;ij  vc;ij ri DrDz

30

ij;/>0

The maximum error in liquid velocity is measured as,


q
2
2
Emax

max
uf;ij  uc;ij vf;ij  vc;ij
Liquid

31

ij;/>0

6.1. Parasitic currents


In this section we test our implementation of surface tension for the problem of a static two-dimensional
(2d) drop with diameter D. We assume the density ratio and viscosity ratio are both one for this problem. The
exact solution for such a problem is that the velocity u is identically zero. If we scale the NavierStokes equations by the time scale T = Dl/r, and by the velocity scale U = r/l, then the non-dimensionalized Navier
Stokes equations become,
Du
rp Oh2 Du  Oh2 jrH
Dt
where the Ohnesorge number Oh is dened as,
l
Oh p
rqD
We investigate the maximum velocity of our numerical method for varying grid resolutions at the dimensionless time t = 250. The dimensions of our computational grid are 5/2 5/2 with periodic boundary conditions
at the left and right boundaries and reecting boundary conditions at the top and bottom boundaries. A drop
with unit diameter is initially located at the center of our domain (5/4,5/4). Our tolerance for the pressure solver and viscous solver is 1.0E  12 (the error is measured as an absolute error and is the L2 norm of the residual). In Table 1 we display results of our grid renement study for 1/Oh2 = 12,000. Our results indicate at least
second-order convergence. These results are comparable to those in [35] where a front tracking method was
used to represent the interface. Our results are also comparable to recent work by [18] in which a height fraction approach for surface tension was tested.
6.2. Surface tension driven (zero gravity) drop oscillations
In this section, we perform a grid renement study for the problem of surface tension driven drop oscillations. In the previous example with parasitic currents, the density ratio was 1:1 and the viscosity ratio was 1:1;
in this example, the density ratio is 1000:1 and the viscosity ratio is 1000:1.
According to the linearized results derived by Lamb [27, Section 275], the position of the drop interface is
Rh; t a P n cosh sinxn t p=2
where
x2n r

nn  1n 1n 2
a3 ql n 1 qg n

Table 1
Convergence study for static droplet with surface tension (parasitic currents test)
Dx

Maximum velocity

2.5/16
2.5/32
2.5/64

7.3E  4
4.5E  6
5.5E  8

Maximum velocity at t = 250 is shown. Oh2 = 1/12000.

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

487

and Pn is the Legendre polynomial of order n. h runs between 0 and 2p, where h = 0 corresponds to r = 0 and
z = a. If viscosity is present, Lamb [27, Section 355] found that the amplitude is proportional to et/s, where
s

a2 qL
lL 2n 1n  1

We compute the evolution of a drop with a = 1, g = 0, lL = 1/50, lL/lG = 1000, r = 1/2, qL = 1 and qL/qG =
1000. The initial interface is given by R(h,0), with  = 0.05 and n = 2. With these parameters we nd x2 = 2.0
and s = 5.0. The uid domain is X = {(r, z)|0 6 r 6 1.5 and 0 6 z 6 1.5} and we compute on grid sizes ranging
from 32 32 to 128 128. The time step for each respective grid size ranges from 0.0007 to 0.000175. Symmetric boundary conditions are imposed at r = 0 and z = 0.
In Table 2, we display the relative error between succeeding resolutions for the minor amplitude RDx(0, t) of
the droplet. The average error Eavg
Amplitude is given by
Z 3:5
Eavg
jRDx 0; t  R2Dx 0; tj dt
Amplitude 
0

and the maximum amplitude error Emax


Amplitude is given by
Emax
Amplitude  max jRDx 0; t  R2Dx 0; tj
06t63:5

In Fig. 9, we plot the minor amplitude versus time for the three dierent grid resolutions.

Table 2
Convergence study for zero gravity drop oscillations r = 1/2, lL = 1/50, lL/lG = 1000, qL/qG = 1000 and a = 2
Dr

Eavg
Amplitude

Emax
Amplitude

3/64
3/128
3/256

N/A
0.00076
0.00021

N/A
0.00172
0.00057

-0.95
"major32"
"major64"
"major128"

-0.96
-0.97

amplitude

-0.98
-0.99
-1
-1.01
-1.02
-1.03
-1.04
-1.05

0.5

1.5

2.5

3.5

time

Fig. 9. Perturbation in minor amplitude for zero gravity drop oscillations (two-phase sharp interface method). lL = 1/50, c = 1/2, density
ratio 1000:1, viscosity ratio 1000:1.

488

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

6.3. Standing wave problem


For the standing wave problem, the free surface at t = 0 is described by the equation
y 1=4  cos2px
where  = 0.025. The gravitational force is g = 2p. We assume inviscid ow, lL = lG = 0, and the density ratio
is 1000, qL = 1, qL/qG = 1000. The computational domain is a 1/2 by 1/2 box with symmetric boundary conditions at x = 0 and x = 1/2 and solid wall boundary conditions at y = 0. In Fig. 10 we compare the amplitude
(at x = 0) for 4 dierent grid resolutions: Dx = 1/64, Dx = 1/128, Dx = 1/256 and Dx = 1/512. The timestep
for each case is Dt = 0.02, Dt = 0.01, Dt = 0.005 and Dt = 0.0025.
In Table 3, we show the relative error between the 4 graphs (0 6 t 6 10). In Table 4, we provide the percent
error for the maximum mass uctuation for the time interval 0 6 t 6 10,
max 100

06t610

jmasst  mass0j
mass0

In Fig. 11, we compare our proposed two-phase sharp interface method to the corresponding onephase method described in Section A.1. They are almost identical, which is expected since our two-phase
sharp interface approach becomes the one-phase approach in the limit of zero gas density qG and zero gas viscosity lG. Also in the same gure, we study the dierence between our sharp interface approach with/without

0.04

"major32"
"major64"
"major128"
"major256"

0.03

amplitude

0.02

0.01

-0.01

-0.02

-0.03

10

time

Fig. 10. Amplitude for inviscid standing wave problem. Density ratio 1000:1 (two-phase sharp interface method).

Table 3
Convergence study: relative error between coarse grid computations with cell size Dxcoarse and ne grid computations with cell size Dxne
for amplitude at x = 0 for standing wave problem
Dxcoarse

Dxne

Maximum error

Average error

1/64
1/128
1/256

1/128
1/256
1/512

2.4E  3
6.5E  4
2.8E  4

6.2E  4
1.5E  4
4.9E  5

Relative error measured for the period 0 6 t 6 10. The physical domain size is 1/2 1/2. Dx is the mesh spacing which is 2n1x where nx is the
number of cells in the x direction. For all our tests, Dx = Dy.

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

489

Table 4
Convergence study: maximum mass uctuation error measured as a percent of the initial mass
Dx

Mass error (%)

1/32
1/64
1/128
1/256

0.078
0.030
0.015
0.007

Mass error measured for the period 0 6 t 6 10. The physical domain size is 1/2 1/2. Dx is the mesh spacing which is
number of cells in the x direction. For all our tests, Dx = Dy.

0.04

1
2nx

where nx is the

"major256"
"major256_singlephase"
"major256_ghostfluid"

0.03

amplitude

0.02

0.01

-0.01

-0.02

-0.03

10

time

Fig. 11. Comparison of two-phase sharp interface method with single-phase method and semi-implicit ghost-uid method. Density
ratio 1000:1.

liquid velocity extrapolation (Step 6 in Section 4). Without liquid velocity extrapolation (a.k.a. the semi-implicit ghost-uid approach), the results do not converge nearly as rapidly as with velocity extrapolation. The no
extrapolation results with Dx = 1/512 are more poorly resolved than the Dx = 1/64 results corresponding to
our sharp-interface approach with liquid velocity extrapolation.
We remark that in (3), we see that the order of accuracy is 1.6 on the nest resolution grids. The order is not
2 since our method is designed to approach a second-order method as qG approaches zero. In this test, we
believe that the error is so small, that the value of qG is big enough to make itself the dominant contribution
to the error. One can also look at qG as being analogous to the cuto used for h in the second-order discretization of the poisson equation on irregular domains [19].
6.4. Traveling wave problem
In [56], experiments were conducted in which traveling waves were generated from wind. In this section we
investigate the performance of our numerical algorithm for simulating traveling waves in the presence of wind.
We shall validate our algorithm by way of a grid renement test. We shall also compare results of our new
algorithm to those results produced by our semi-implicit ghost-uid method.
According to [56], a wind velocity of U = 5 m/s will generate traveling waves with a wavelength k = 15 cm,
a phase velocity C = 50 cm/s, a wave period P = 0.3 s, and a trough to peak wave height H = 1 cm. Also, for a
wind speed of U = 5 m/s, the roughness length is z0a = 0.3 cm and the friction velocity is u*a = 30 cm/s.

490

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

We initialized our computational domain as a 15 30 cm rectangular box with the initial position of the
water surface given by,
y surface x 15:0

H
cos2px=k
2

We shall assume periodic boundary conditions on the left and right walls, and free-slip boundary conditions
on the upper and lower walls.
The initial velocity in the water is derived using a similar numerical procedure as found in [59]. We compute
a stream function w which is dened in the whole computational domain. In the calculation of w, we assume
the initial vorticity is zero everywhere except on the interface. The vortex sheet strength at the airwater interface is given by,
C H x coskx;
where k = 2p/k is the wave number and x and k satisfy the following linear dispersion relation:
rk 3
qL

x2 gk

32

We have ignored the gas density qG and the water depth (15 cm) in (32) since these values have a negligible
eect on x. In our computations, we used the actual physical properties for air and water: qL = 1.0 g/cm3,
qG = 0.001229 g/cm3, lL = 0.0089 g/(cm s), lG = 1.73e  4 g/(cm s), g = 980.0 cm/s2, and r = 72.8 dyne/cm.
Given these properties, we have x = 20.39. We remark that the linear dispersion relation predicts a period
of P = 2p/x = 0.31 which is very close to the experimental values reported by [56].
Given the vortex sheet strength, we solve for the stream function w using the following equation:
wxx wyy CjrH /j

33

|$H| is discretized as,


s

2 
2
H i1=2;j  H i1=2;j
H i;j1=2  H i;j1=2

Dx
Dy
where,

H i1=2;j

1 /ij P 0 or /i1;j P 0
0 otherwise

Once w is found, we have:


wi;j1  wi;j1
2Dy
wi1;j  wi1;j
vij 
2Dx
uij

34
35

The boundary conditions for w in (33) are homogeneous Dirichlet conditions at the top and bottom of the
computational domain, and periodic boundary conditions on the left and right sides.
In [59], the velocity in the air as well as in the water was given by (34) and (35). In our test, we shall initialize
the air velocity to have the characteristic logarithmic wind prole given by,
(
0
y < y surface x z0a
ux; y ua
yy surface x
log z0a otherwise
K
vx; y 0
where K = 0.4 is von Karmons constant, z0a = 0.3 cm is the roughness length, and u*a = 30.0 cm/s is the friction velocity.

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

491

Given the cell centered initial velocity in the water and air, we interpolate these respective velocity elds
from cell centers to cell faces and then we initialize the face centered velocity V as,
( L
U i1=2;j /ij P 0 or /i1;j P 0
V i1=2;j
UG
otherwise
i1=2;j
The initial velocity should be divergence free so we project V as described in Section 5.5 in order to insure a
discretely divergence free initial velocity eld. After the projection step, we initialize the liquid and gas velocity
with the projected velocity U and then we extend the liquid velocity into the gas in order to construct UL. In
Fig. 12 we plot the initial velocity elds UL and U.
In Fig. 13, we compare the amplitude (at x = 0) versus time for three dierent grid resolutions: Dx = 15/32,
Dx = 15/64 and Dx = 15/128. The timestep for each case is Dt = 0.0008, Dt = 0.0004, and Dt = 0.0002. In
Table 5, we show the relative error between the 3 graphs (0 6 t 6 1). In Fig. 14, we plot the amplitude for
our sharp interface without liquid velocity extrapolation (Step 6 in Section 4). Without liquid velocity extrapolation (a.k.a. the semi-implicit ghost-uid approach), the results do not converge nearly as rapidly as with
velocity extrapolation. The no extrapolation results with Dx = 15/128 are more poorly resolved than the
Dx = 15/64 results corresponding to our sharp-interface approach with liquid velocity extrapolation.

Fig. 12. Initial velocity eld for wind driven wave problem. Left: initial liquid velocity UL derived from stream function. Right: initial
combined liquid/gas velocity U. Wind velocity in the gas has logarithmic prole. Grid resolution is 128 256.

492

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505


0.8

"32x64"
"64x128"
"128x256"

0.6

amplitude

0.4

0.2

-0.2

-0.4

-0.6

0.2

0.4

0.6

0.8

time

Fig. 13. Amplitude for traveling wave problem with wind. Density ratio 813:1 (two-phase sharp interface method).
Table 5
Convergence study: relative error between coarse grid computations with cell size Dxcoarse and ne grid computations with cell size Dxne
for amplitude at x = 0 for traveling wave problem with wind
Dxcoarse

Dxne

Maximum error

Average error

15/32
15/64

15/64
15/128

0.122
0.057

0.031
0.014

Relative error measured for the period 0 6 t 6 1. The physical domain size is 15 30. Dx is the mesh spacing which is
number of cells in the x direction. For all our tests, Dx = Dy.

0.8

15
nx

where nx is the

"32x64_no_extrapolation"
"64x128_no_extrapolation"
"128x256_no_extrapolation"

0.6

amplitude

0.4

0.2

-0.2

-0.4

-0.6

0.2

0.4

0.6

0.8

time

Fig. 14. Amplitude for traveling wave problem with wind. Density ratio 813:1. Velocity extrapolation is disabled (semi-implicit ghostuid method).

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

493

Remarks:
 We measured rst-order accuracy using our sharp-interface method (with velocity extrapolation). We
attribute this to how we obtained our initial velocity eld in the liquid. The discretization to the right
hand side of (33) is a low order approximation to the delta function; nonetheless, we see signicant
improvement in our calculations with velocity extrapolation, as opposed to without. Without velocity
extrapolation, we do not see any convergence for the grid sizes used.
 The computations in [59] (using a continuum approach) were limited to a wave Reynolds number of
around 150 and a density ratio of 100:1. Also, wind was not taken into account in their computations. In
the results we present here, using the actual physical properties of air and water, the wave Reynolds
L
number is Re q2plCkL 12875 and the density ratio is 813:1.
6.5. Capillary instability
In this section, we test our sharp-interface approach on the classical Rayleigh capillary instability problem in which a slightly perturbed cylindrical column of liquid is driven to break up into droplets by surface
tension (capillary) eects. In this test problem we use parameters that are comparable to those found in
[50].
We consider an initially perturbed cylindrical column of water in air. The shape of the initial interface is
rz r0  cos2pz=k

36

We compute on a 3d-axisymmetric domain X = {(r, z)|0 6 r 6 k/4 and 0 6 z 6 k/2}. Symmetric boundary conditions are enforced at r = 0, z = 0 and z = k/2. Outow (pressure equals zero) boundary conditions are enforced at r = k/4. The relevant dimensional parameters for this test problem are r0 = 6.52 lm,  = 1.3 lm,
k = 60 lm, lL = 1.138 102 g/(cm s), lG = 1.77 104 g/(cm s), qL = 1.0 g/cm3, qG = 0.001225 g/cm3, and
r = 72.8 dynes/cm. In our computations we use the following dimensionless parameters: the Reynolds number
Re = qLLU/lL = 7.5, the Weber number We = qLLU2/r = 1.0, L = 1 lm, U = 8.53 m/s and the density and
viscosity ratios are 816 and 64, respectively.
In Fig. 15, we display the results of our computations for the capillary jet as it breaks up. In Table 6, we
measure the relative errors for the interface and velocity eld for grid resolutions ranging from 16 32 to
64 128. The time step ranged from Dt = 0.04 to Dt = 0.01.
As shown in Table 6, we obtain about rst-order accuracy for the solution in the liquid. We attribute our
low order accuracy to how we discretize the viscosity term,
1
r  2lD;
q

37

at the interface. Suppose lG = 0 and the zero level set crosses between cells (i, j) (/ij < 0) and (i + 1, j)
(/i+1, j P 0). In this case the values for l and q jump from lG to lL and from qG to qL abruptly where the
level set function changes sign; i.e. our discretization of l and q in Eq. (37) does not incorporate specic information about the location of the zero level set in between cells (i, j) and (i + 1, j), except that the zero level set is
somewhere between these two cells. We remark that, although we observe rst-order accuracy using our sharp
interface approach, our errors are considerably smaller than those presented in [50]. We also get comparable
results when calculating the break-up of a liquid jet using our single-phase method (Section A.1, see Table 8
and Fig. 16).
If we reduce the viscosity further, i.e. set the Reynolds number Re = 200, then we get much closer to second-order convergence using our sharp interface approach, as illustrated in Table 7.
6.6. Bubble dynamics
In this section, we compute the steady state shapes of a gas bubble rising in a viscous Newtonian liquid. For
comparison, we use the experimental results found in [7,25] and computational results in [39].

494

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

t=110.0

t=100.0

t=80.0

t=40.0
Fig. 15. Capillary instability. Two-phase sharp interface method. qL/qG = 816, lL/lG = 64. Grid resolution is 64 128.

Table 6
Convergence study for the Rayleigh capillary instability problem using the two-phase sharp interface method
Grid

Einterface

Eavg
Liquid

Emax
Liquid

Eavg
vapor

16 32
32 64
64 128

N/A
14.4
7.9

N/A
8.0
4.5

N/A
0.012
0.009

N/A
24.8
11.6

Elapsed time is t = 80. The viscosity and density ratios are lL/lG = 64 and qL/qG = 816, respectively. The Reynolds number is 7.5.

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

495

Table 7
Convergence study for the Rayleigh capillary instability problem using the two-phase sharp interface method
Grid

Einterface

Eavg
Liquid

Emax
Liquid

Eavg
vapor

16 32
32 64
64 128

N/A
4.2
0.9

N/A
3.2
1.1

N/A
0.013
0.004

N/A
32.8
11.1

Elapsed time is t = 80. The viscosity and density ratios are lL/lG = 64 and qL/qG = 816, respectively. The Reynolds number is 200.

Table 8
Convergence study for the Rayleigh capillary instability problem using the single-phase sharp interface method
Grid

Einterface

Eavg
Liquid

Emax
Liquid

16 32
32 64
64 128

N/A
13.6
7.5

N/A
7.8
4.3

N/A
0.012
0.011

Elapsed time is t = 80. The Reynolds number is 7.5.

As in [7,25], we will present our computational results in terms of the following dimensionless groups. The
Reynolds number Re, the Eotvos number Eo, and the Morton number Mo are dened as follows:
Re

qLU
gL

Eo

gL2 U
r

Mo

gg4L
qr3

38

where q is the liquid density, L is the bubble diameter, U is a characteristic velocity, gL is the liquid viscosity, r
is the surface tension, and g is the acceleration of gravity.
Another set of useful dimensionless numbers, although not independent of those in (38), are the Weber
number We, the Froude number Fr, and the drag coecient CD:
We

qLU 2
r

Fr

U2
gL

CD

4qgL2
3gL U

In all of our bubble calculations, we use adaptive mesh renement[46] with a base coarse grid of 24 72 grid
cells and three levels of adaptivity. The computational domain size was 2.0 6.0. Our computations use 3d-axisymmetric rz coordinates. A comparison of computed terminal bubble rise velocity versus previous computational and experimental results are reported in Table 9. A comparison of computed terminal bubble shapes
versus previous computational and experimental results are reported in Fig. 17. Our comparisons include
oblate ellipsoidal cap bubbles studied by [7] (Eo = 243, Mo = 266, and Re = 7.77 for bubble Fig. 2(d) and
Eo = 116, Mo = 5.51, and Re = 13.3 for bubble Fig. 3(d)), spherical cap bubbles studied by Hnat and Buckr
master [25] (Re = 19.4, Mo = 0.065, and C = 4.95, where C m2 =g
1=3 ), and a disk-bubble studied by Ryskin
and Leal [39] (R = 100 and We = 10).
Finally, we remark that for these bubble rise test problems, the semi-implicit ghost-uid approach (our
two-phase approach with velocity extrapolation disabled) produces results comparable with our two-phase
approach. Results for the semi-implicit ghost-uid approach are shown in Table 10 and Fig. 18.
6.6.1. Full 3d bubble dynamics
As a validation of our sharp interface method in 3 dimensions, we compute bubble motion in 3d-Cartesian
coordinates (x, y, and z) and compare our results to the corresponding 3d-axisymmetric computations. The
dimensions of the computational domain was 4 4 6. We computed 3d bubble motion on an adaptive grid
with base coarse grid size of 16 16 24 and 3 additional levels of adaptivity. In Fig. 19 we show the computed bubble shape in which we used the same physical properties as the D = 12.15 case in Hnat and Buckmasters paper [25]. The experimental rise speed (in terms of the Re number) is 19.4 and our computed rise
speed is 19.5. In Fig. 20 we show the computed bubble shape in which we used the same physical properties
as in Fig. 3(d) of Bhaga and Webers paper [7]. The experimental rise speed (in terms of the Re number) is 13.3
and our computed rise speed is 13.6.

496

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

t=110.0

t=100.0

t=80.0

t=40.0
Fig. 16. Capillary instability. Single-phase sharp interface method. qL/qG = 816, lL/lG = 64. Grid resolution is 64 128.

Table 9
Comparison of computed terminal bubble rise speed (in terms of the Re number) compared with experiments (Bhaga and Weber,
Buckmaster) and compared with previous calculations (Ryskin and Leal)
Case

Sharp interface method

Experiment/previous result

Fig. 2d (Bhaga and Weber)


Fig. 3d (Bhaga and Weber)
Ryskin and Leal (Re = 100, We = 10)
Buckmaster (D = 12.15)

8.3
14.1
97.5
19.8

7.8
13.3
100
19.4

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

497

Fig. 17. Comparison of our numerical results (two-phase sharp interface method) with experimental/benchmark results. Upper left: Bhaga
and Weber (Fig. 2, bubble (d)). Upper right: Bhaga and Weber (Fig. 3, bubble (d)). Lower left: Hnat and Buckmaster. Lower right: Ryskin
and Leal.
Table 10
Comparison of computed terminal bubble rise speed (in terms of the Re number) using the semi-implicit ghost-uid sharp interface
method compared with experiments (Bhaga and Weber, Buckmaster) and compared with previous calculations (Ryskin and Leal)
Case

Semi-implicit ghost-uid

Experiment/previous result

Fig. 2d (Bhaga and Weber)


Fig. 3d (Bhaga and Weber)
Ryskin and Leal (Re = 100, We = 10)
Buckmaster (D = 12.15)

8.1
13.7
97.6
19.7

7.8
13.3
100
19.4

6.7. Bubble formation


In this section we compute the formation of bubbles caused by the injection of air into a container of liquid.
Our computations use 3d-axisymmetric rz coordinates. We enforce inow boundary conditions at the bottom
of the domain (z = 0),
rp  n 0

uinflow
U n
0

r < rnozzle
otherwise

498

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

Fig. 18. Comparison of our numerical results (two-phase semi-implicit ghost-uid method) with experimental/benchmark results. Upper
left: Bhaga and Weber (Fig. 2, bubble (d)). Upper right: Bhaga and Weber (Fig. 3, bubble (d)). Lower left: Hnat and Buckmaster. Lower
right: Ryskin and Leal.

Fig. 19. Full 3d computations of a rising gas bubble in liquid. Physical properties correspond to the D = 12.15 case in Hnat and
Buckmaster. Left: side. Right: bottom.

Symmetry boundary conditions are given at r = 0, free-slip conditions at r = rhigh, and outow conditions at
the top of the domain (z = zhigh):
p0

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

499

Fig. 20. Full 3d computations of a rising gas bubble in liquid. Physical properties correspond to Fig. 3(d) case in Bhaga and Weber. Left:
side. Right: bottom.

Fig. 21. Bubble formation computed using two-phase sharp interface method. Nozzle radius 8.5E  4m. Inow velocity 0.44 m/s. Density
ratio 1015:1, Viscosity ratio 6923:1.

500

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

We compare results of our two-phase sharp interface method with experimental results reported by Helsby
and Tuson [23]. Our target is Fig. 1(e) in [23]. This corresponds with a nozzle radius of 8.5E  4m and an
inow velocity of 0.44 m/s. Based on the physical properties of the case-e system, one has the Reynolds number equal to 3.6, the Weber number equal to 3.06, the density ratio equal to 1015:1 and the viscosity ratio
equal to 6923:1. We used Adaptive mesh renement [46] to compute the solutions for the bubble formation
problem with a base coarse grid of 32 96 grid cells and three levels of adaptivity. There were 16 ne grid
cells spanning the nozzle radius. In Fig. 21 we illustrate our computational results. The bubble diameters
for the 2nd and 3rd bubbles were 4.85E  3m and 4.90E  3m, respectively, which is in good agreement with
the experimental result 4.99E  3m.
7. Conclusions
A sharp interface method for two-phase ows has been developed. Our method has been designed to
reduce to a single-phase approach in the limiting case of zero gas density and zero gas viscosity. Also,
a new cell-centered semi-implicit treatment for the viscous terms has been developed which enables us to
bypass the viscous time step constraint while treating the viscosity jump as sharp. For problems with
a thin free-surface boundary layer, our results are superior to the semi-implicit ghost-uid method.
For problems in which the Reynolds number is large in the liquid, our results demonstrate second-order
accuracy for the liquid solution of two-phase incompressible ows. For problems in which viscous eects
are dominant, both our two-phase and one-phase sharp interface approaches become rst-order accurate. In fact, the errors of all three approaches, (1) our proposed sharp interface method, (2) our semiimplicit ghost-uid method, and (3) our single-phase method are all comparable to each other when
viscous eects are suciently present. When viscous eects are weak, then our sharp interface approach
gives higher accuracy than our semi-implicit ghost-uid approach. This is expected, since it is for this
class of problems that the solutions admitted from a ghost-uid approach (which assumes continuity of
the tangential velocity) diverge from our sharp-interface approach (and diverge from a one-phase
approach). The improved accuracy over conventional rst-order continuum approaches and ghostuid approaches allows us to resolve computations using a coarse mesh where otherwise a ne mesh is
required. As demonstrated in our bubble formation test, our new method can reliably handle complex
interfacial geometries.
Acknowledgments
We thank D. Kikuchi and S. Yamaguchi for their help in preparing this manuscript.
Appendix A
A.1. One-phase algorithm
The one-phase algorithm addresses a two-phase ow problem in which the liquid is assumed to behave
incompressibly, and the pressure in the gas is spatially constant [45,17,11].
In the liquid we have,
DU
r  pI 2lD qg^z
Dt
rU 0
q

where U is the velocity vector, q is the density, p is the pressure, l is the coecient of viscosity, g is the
gravity, I is the unit tensor, ^z is the unit vector in the vertical direction, and D is the deformation tensor
dened by
D

rU rU
2

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

501

In the vapor, we assume p(t) is constant in space. The vapor viscosity lG and density qG are assumed to be
zero. The free surface boundary conditions are enforced by specifying the following pressure boundary condition at the free surface:
px; t pvapor t  rj 2lL DL  n  n

39

where j is the local mean curvature, lL is the liquid viscosity, and DL is the rate of deformation for the liquid.
If one denes the interface C as the zero level set of a smooth level set function, /, then the resulting equations are written as:
DU
r  pI 2lD qg^z  rj  pvapor trH
Dt
rU 0
D/
0 q qL H / l lL H /
Dt

40

41

where j(/) and H(/) are dened by Eqs. (3) and (4), respectively.

curvature "A"

curvature "B"

Curvature "C"

Fig. 22. The volume fractions in the following 3 7 stencil are used to approximate curvature A to second-order accuracy. In order to
compute curvature B to second-order accuracy, one must linearly interpolate between curvature A and curvature C.

502

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

Boundary conditions must be specied in the vapor (/ < 0). The boundary conditions are p = 0 and
U U liquid
extrapolate . In the pressure projection step, the density is expressed in terms of the height fraction (see
Eq. (28) except replace qG with zero). The discretization of the pressure projection step is second-order accurate (see [19]). As in [19,45], we prescribe a cuto for the height fraction hi+1/2, j (see Eq. (28)) which is 0.001.
Further details for the discretization of (40) thru (41) are given in [45].
A.2. Curvature discretization
The curvature on the free surface is computed to second-order accuracy directly from the volume fractions
[22]. Previous work in this area include that by Chorin [13], Poo and Ashgriz [34], Aleinov and Puckett [1],
Williams et al. [55] and more recently, using PROST, Renardy et al. [37]. The method we use here is explicit,
localized, and can be shown thru Taylor series expansion to be second-order accurate for rz or 3d coordinate
systems. The method is based on reconstructing the height function directly from the volume fractions [22].
Without loss of generality, we assume that the free surface is oriented more horizontal than vertical. The orientation of the free surface is determined from the level set function since n = $//|$/|. A 3 7 stencil of volume fractions is constructed about cell (i,j) (see Fig. 22). The 3 vertical
sums, F i0 , i 0 = i  1, i, i + 1 correspond
R xi1=2
1
to the integrals of the height function h(x) (see Fig. 23); i.e. F i Dx xi1=2 hx dx Cj. It can be shown that
(Fi+1  Fi1)/(2Dx) is a second-order approximation to h 0 (xi) and that (Fi+1  2Fi + Fi1)/Dx2 is a secondorder approximation to h00 (xi). A slightly more complicated procedure is used in axisymmetric coordinate systems; the height function h(r) is assumed to have the form ar2 + br + c. The integral of rh(r) is related with F i0 ,
i 0 = i  1, i, i + 1 in order to solve for the 3 unknowns a, b and c. For vertically oriented interfaces in axisymmetric coordinateR systems, the F j0 represent the integrals of the square of the height function h(z) (up to a conz
1
stant): F j0 Dz
p z j00 1=2 hz2 dz Ci. In other words, (Fj+1  Fj1)/(2Dz) is a second-order approximation
j 1=2
to dh(z)2/dz and (Fj+1  2Fj + Fj1)/Dz2 is a second-order approximation to d2(h(z)2)/dz2. The resulting
cur777vature is obtained directly from the height function (whether it be h(r), h(z) or h(x, y)).
This procedure for nding curvature will return a second-order approximation to the curvature on the
interface passing thru cell (i, j) located at x = (i + 1/2)Dx (horizontal orientation) or y = (j + 1/2)Dy (vertical

(i1,j+3)

(i+1,j+3)

h(x)

(i1,j3)

(i+1,j3)

(i1/2,j7/2) (i+1/2,j7/2)
Fig. 23. Stencil for calculating the curvature in cell (i, j) when the level set function changes sign between cells (i, j) and (i, j + 1). The shaded
P
area corresponds to the
vertical sum of the volume fractions, DxDy j3
J j3 F i;J , and the shaded area also corresponds to the integral of the
R xi1=2
height function h(x), xi1=2
hx dx Cj.

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

503

Table 11
Convergence study for computing curvatures from volume fractions of a unit sphere in axisymmetric geometry
Dx

Maximum error

Average error

1/16
1/32
1/64

0.0104
0.0024
0.0006

0.0037
0.0009
0.0002

The physical domain size is 2 4. Dx is the mesh spacing which is 2/nx, where nx is the number of cells in the x direction. For all of our
tests, Dx = Dy.

Table 12
Convergence study for computing curvatures from volume fractions of a unit sphere in three-dimensional geometry
Dx

Maximum error

Average error

1/8
1/16
1/32

0.094
0.050
0.010

0.0125
0.0036
0.0009

The physical domain size is 4 4 4. Dx is the mesh spacing which is 4/nx, where nx is the number of cells in the x direction. For all of our
tests, Dx = Dy = Dz.

orientation). In order to nd jI(F) to second-order accuracy (17), we have two dierent cases when the level set
function changes sign between cells (i, j) and (i + 1, j): (1) the interface is orientated vertically, in which case

h < 1=2
jij
jI
ji1;j otherwise
or (2) the interface is orientated horizontally, in which case
jI 1  hjij hji1;j
In Tables 11 and 12, we display the average error and maximum error for the case of a sphere in axisymmetric and three-dimensional coordinate systems, respectively.
References
[1] I. Aleinov, E.G. Puckett, Computing surface tension with high-order kernels, in: Proceedings of the 6th International Symposium on
Computational Fluid Dynamics, Lake Tahoe, CA, 1995.
[2] A.S. Almgren, J.B. Bell, P. Colella, L.H. Howell, M. Welcome, A conservative adaptive projection method for the variable density
incompressible NavierStokes equations, J. Comput. Phys. 142 (1998) 146.
[3] E. Aulisa, S. Manservisi, R. Scardovelli, A mixed markers and volume-of-uid method for the reconstruction and advection of
interfaces in two-phase and free-boundary ows, J. Comp. Phys. 188 (2003) 611.
[4] E. Aulisa, S. Manservisi, R. Scardovelli, A surface marker algorithm coupled to an area-preserving marker redistribution method for
three-dimensional interface tracking, J. Comp. Phys. 197 (2) (2004) 555584.
[5] J.B. Bell, P. Colella, H.M. Glaz, A second-order projection method for the incompressible NavierStokes equations, J. Comput. Phys.
85 (December) (1989) 257283.
[6] J.B. Bell, D.L. Marcus, A second-order projection method for variable-density ows, J. Comput. Phys. 101 (1992) 334348.
[7] D. Bhaga, M.E. Weber, Bubbles in viscous liquids: shapes, wakes and velocities, J. Fluid Mech. 105 (1981) 6185.
[8] A. Bourlioux, A coupled level-set volume-of-uid algorithm for tracking material interfaces, in: Proceedings of the 6th International
Symposium on Computational Fluid Dynamics, Lake Tahoe, CA, 1995.
[9] J.U. Brackbill, D.B. Kothe, C. Zemach, A continuum method for modeling surface tension, J. Comput. Phys. 100 (1992) 335353.
[10] S. Bradford, N.D. Katopodes, Hydrodynamics of turbid underows. Part II: aggradation, avulsion and channelization, Journal of
Hydraulic Engineering, ASCE 125 (10) (1999) 10161028.
[11] A. Caboussat, M. Picasso, J. Rappaz, Numerical simulation of free surface incompressible liquid ows surrounded by compressible
gas, J. Comput. Phys. 203 (2) (2005) 626649.
[12] Y.C. Chang, T.Y. Hou, B. Merriman, S. Osher, Eulerian capturing methods based on a level set formulation for incompressible uid
interfaces, J. Comput. Phys. 124 (1996) 449464.
[13] A.J. Chorin, Curvature and solidication, J. Comput. Phys. 57 (1985) 472490.
[14] T.J. Craft, J.W. Kidger, B.E. Launder, Three dimensional modelling of turbulent free-surface jets, in: W. Rodi, D. Laurence (Eds.),
Engineering Turbulence Modelling and Measurements, vol. 4, Elsevier, Amsterdam, 1999, pp. 7382.

504

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

[15] D.G. Dommermuth, M. Gharib, H. Huang, G.E. Innis, P. Maheo, E.A. Novikov, J.C. Talcott, D.C. Wyatt, Turbulent free-surface
ows: a comparison between numerical simulations and experimental measurements, in: Proceedings of the 21st Symposium on Naval
Hydro, Trondheim, 1996, pp. 249265.
[16] D. Enright, R. Fedkiw, J. Ferziger, I. Mitchell, A hybrid particle level set method for improved interface capturing, J. Comp. Phys.
183 (1) (2002) 83116.
[17] D. Enright, S. Marschner, R. Fedkiw, Animation and rendering of complex water surfaces, in: SIGGRAPH 2002, volume ACM TOG
21, 2002, pp. 736744.
[18] M. Francois, S. Cummins, E. Dendy, D. Kothe, J. Sicilian, M. Williams, A balanced-force algorithm for continuous and sharp
interfacial surface tension models within a volume tracking framework, J. Comput. Phys. 213 (1) (2006) 141173.
[19] F. Gibou, R. Fedkiw, L.-T. Cheng, M. Kang, A second order accurate symmetric discretization of the poisson equation on irregular
domains, J. Comput. Phys. 176 (2002) 205227.
[20] D. Gueyer, J. Li, A. Nadim, R. Scardovelli, S. Zaleski, Volume of uid interface tracking with smoothed surface stress methods for
three-dimensional ows, J. Comput. Phys. 152 (1999) 423456.
[21] B. Helenbrook, L. Martinelli, C. Law, A numerical method for solving incompressible ow problems with a surface of discontinuity,
Journal of Computational Physics 148 (1999) 366396.
[22] J. Helmsen, P. Colella, E.G. Puckett, Non-convex prole evolution in two dimensions using volume of uids, LBNL Technical Report
LBNL-40693, Lawrence Berkeley National Laboratory, 1997.
[23] F.W. Helsby, K.R. Tuson, Behaviour of air bubbles in aqueous solutions, Research 8 (1955) 270.
[24] S. Hieber, P. Koumoutsakos, A Lagrangian particle level set method, Journal of Computational Physics 210 (1) (2005) 342367.
[25] J.G. Hnat, J.D. Buckmaster, Spherical cap bubbles and skirt formation, Physics of Fluids 19 (2) (1976) 182194.
[26] M. Kang, R. Fedkiw, X.-D. Liu, A boundary condition capturing method for multiphase incompressible ow, J. Sci. Comput. 15
(2000) 323360.
[27] H. Lamb, Hydrodynamics, Dover Publications, New York, 1932.
[28] H. Liu, S. Krishnan, S. Marella, H.S. Udaykumar, Sharp interface cartesian grid method II: a technique for simulating droplet
interactions with surfaces of arbitrary shape, J. Computational Physics 210 (1) (2005) 3254.
[29] Daniel Lorstad, Laszlo Fuchs, High-order surface tension VOF-model for 3D bubble ows with high density ratio, Journal of
Computational Physics 200 (1) (2004) 153176.
[30] E. Marchandise, J.-F. Remacle, N. Chevaugeon, A quadrature-free discontinuous Galerkin method for the level set equation, Journal
of Computational Physics 212 (1) (2006) 338357.
[31] M.L. Minion, On the stability of Godunov-projection methods for incompressible ow, J. Comput. Phys. 123 (2) (1996) 435449.
[32] D. Nguyen, R. Fedkiw, M. Kang, A boundary condition capturing method for incompressible ame discontinuities, J. Comput. Phys.
172 (2001) 7198.
[33] J.E. Pilliod, E.G. Puckett, Second-order accurate volume-of-uid algorithms for tracking material interfaces, Journal of
Computational Physics 199 (2) (2004) 465502.
[34] J.Y. Poo, N. Ashgriz, A computational method for determining curvatures, J. Comput. Phys. 84 (1989) 483491.
[35] S. Popinet, S. Zaleski, A front-tracking algorithm for accurate representation of surface tension, International Journal for Numerical
Methods in Fluids 30 (6) (1999) 775793.
[36] H. Rasmussen, A. Bach, O. Hassager, 3D transient free surface viscoelastic ow, in: XIIth International Workshop on Numerical
Methods for Non-Newtonian Flows, Abstracts, June, 2001, Monterray Bay, USA.
[37] Y. Renardy, M. Renardy, Prost: A parabolic reconstruction of surface tension for the volume-of-uid method, J. Comput. Phys. 183
(2) (2002) 400421.
[38] M. Rudman, A volume tracking method for interfacial ows with large density variations, Int. J. Numer. Methods Fluids 28 (1998)
357378.
[39] G. Ryskin, L.G. Leal, Numerical solution of free boundary problems in uid mechanics. Part 2 buoyancy-driven motion of a gas
bubble through a quiescent liquid, J. Fluid Mech. 148 (1984) 1935.
[40] R. Scardovelli, S. Zaleski, Interface reconstruction with least-square t and split EulerianLagrangian advection, Int. J. Numer.
Meth. Fluids 41 (2003) 251274.
[41] Seungwon Shin, Damir Juric, Modeling three-dimensional multiphase ow using a level contour reconstruction method for front
tracking without connectivity, J. Comp. Phys. 180 (2) (2002) 427470.
[42] C.-W. Shu, S. Osher, Ecient implementation of essentially non-oscillatory shock capturing schemes, ii, J. Comput. Phys. 83 (1989)
3278.
[43] J. Strain, Tree methods for moving interfaces, J. Comput. Phys. 151 (2) (1999) 616648.
[44] G. Strang, On the construction and comparison of dierence schemes, SIAM J. Numer. Anal. 5 (1968) 506517.
[45] M. Sussman, A second order coupled levelset and volume of uid method for computing growth and collapse of vapor bubbles,
Journal of Computational Physics 187 (2003) 110136.
[46] M. Sussman, A parallelized, adaptive algorithm for multiphase ows in general geometries, Computers and Structures 83 (2005) 435
444.
[47] M. Sussman, A. Almgren, J. Bell, P. Colella, L. Howell, M. Welcome, An adaptive level set approach for incompressible two-phase
ows, J. Comput. Phys. 148 (1999) 81124.
[48] M. Sussman, D. Dommermuth, The numerical simulation of ship waves using cartesian grid methods, in: Proceedings of the 23rd
Symposium on Naval Hydrodynamics, September, 2000, Val-De-Reuel, France.

M. Sussman et al. / Journal of Computational Physics 221 (2007) 469505

505

[49] M. Sussman, M.Y. Hussaini, A discontinuous spectral element method for the level set equation, J. Scientic Computing 19 (2003)
479500.
[50] M. Sussman, E.G. Puckett, A coupled level set and volume of uid method for computing 3D and axisymmetric incompressible twophase ows, J. Comp. Phys. 162 (2000) 301337.
[51] M. Sussman, P. Smereka, S.J. Osher, A level set approach for computing solutions to incompressible two-phase ow, J. Comput.
Phys. 114 (1994) 146159.
[52] O. Tatebe, The multigrid preconditioned conjugate gradient method, in: 6th Copper Mountain Conference on Multigrid Methods,
Copper Mountain, CO, April 49, 1993.
[53] S.O. Unverdi, G. Tryggvason, A front-tracking method for viscous, incompressible, multi-uid ows, J. Comput. Phys. 100 (1992)
2537.
[54] B. van Leer, Towards the ultimate conservative dierence scheme. V. A second-order sequel to Godunovs method, J. Comput. Phys.
32 (1979) 101136.
[55] M. Williams, D. Kothe, E.G. Puckett, Convergence and accuracy of kernel-based continuum surface tension models, in: Proceedings
of the 13th U.S. National Congress of Applied Mechanics, Gainesville, FL, June 1621, 1998.
[56] J. Wu, Wind-induced drift currents, J. Fluid Mech. 68 (1975) 4970.
[57] B. Yang, A. Prosperetti, A second-order boundary-tted projection method for free-surface ow computations, J. Comput. Phys. 213
(2006) 574590.
[58] X. Yang, A. James, J. Lowengrub, X. Zheng, V. Cristini, An adaptive coupled level-set/volume-of-uid interface capturing method
for unstructured triangular grids, J. Comput. Phys. (in press), doi:10.1016/j.jcp.2006.01.007.
[59] Y. Yang, G. Tryggvason, Dissipation of energy by nite amplitude surface waves, Computers and Fluids 27 (1998) 829845.
[60] T. Ye, W. Shyy, J.N. Chung, A xed-grid, sharp interface method for bubble dynamics and phase change, J. Comp. Phys. 174 (2001)
781815.